The Queue's First Live Evening

Why Care?

Shipping a feature and operating it are different sports. The search-results queue landed earlier tonight ([[2026-07-24_07_Searches-Become-Async-Jobs-The-Queue-Rail-Lands-Fire-Many-Walk-Away-Triage-On-Arrival|ship note]]), and within the hour the operator was firing real concurrent searches at real orgs — which surfaced five things no browser drive had: a missing per-card dismiss, crawls secretly running single-file, result rows overflowing the rail, a resize drag that fought its own hover states, and an API credit balance that kept emptying itself. All five were diagnosed, fixed, verified live, and merged to rebuild/turbo-rsbuild the same night. This entry is the record of that sweep.

What's New?

Per-card × dismiss. The header's "clear done" cleared the whole queue; dismissing ONE search meant expanding it first. Every card now carries a tiny × above its expand caret — one tap. A done card with unreviewed candidates flips the × to a ✓? for four seconds instead (the same conscience as the expanded confirm, compressed).
Crawls run three wide. The crawl consumer awaited each NATS message inside its loop, so concurrent submissions — the queue's whole premise — ran single-file: a batch of five put the last two past the caller's 600s ceiling ("timeout", masquerading as a billing error). The loop now spawns per message under a slot cap (MAX_CONCURRENT_CRAWLS, default 3). Measured: an NYT links crawl that took 4:35–6:39 serialized landed in 1:14 parallel.
Rows fit the rail. Grid children default to min-width: auto and refuse to shrink below their content, so one long nowrap title pushed cards wider than the rail — cutting off the ➕ entirely. Zeroed down the chains; the same disease was then found and cured in the org card's people list.
The flow resize-drag actually resizes. The focused-panel edge drag computed width as distance-from-stage-centre × 2 — correct only when the panel is centred, so first/last flow steps jumped on grab and tracked 2× per pixel; mid-drag, the peek overlay's hover-expand fought the geometry. Now the edge position maps per-arrangement, the edge takes pointer capture, hover is suppressed while resizing, and edges only render on sides that border a peek. Verified: pointer at 50/65/80% of the stage → panel at exactly 50/65/80%.
The burn-rate hole is capped. The crawl's server-side web_search tool carried no max_uses — a crawl of a huge publisher (NYT) searched open-endedly, which is how a same-evening credit top-up drained back to "balance too low" within the hour. Crawls now carry CRAWL_MAX_WEB_SEARCHES (default 8); chat/enrichment calls keep their existing behavior.

Under the Hood

The evening also proved the queue's failure honesty end-to-end: billing 400s landed on cards verbatim with Retry; a workspace restart marked its stranded search failed-with-retry instead of spinning; a dismissed running card kept the crawl billing in the background — which is now [[https://github.com/lossless-group/augment-it/issues/46|gh #46]] (crawls should get the prompt.run.cancel abort treatment so × means what the operator thinks it means).

Merge trail: PR #43 (queue + hardening) merged clean; stacked PR #45 was auto-closed by GitHub when its base branch merged and was deleted — reopened as its named successor #47 and merged. Issues #42 and #44 closed by the merges; #46 and #35 (progress traces) are the queue's open roadmap.

One operating lesson worth keeping: the relevance brief steers crawls well for entities that fit its frame (funders and operators), but gives weak purchase on giant general publishers — for those, hand-adding the two or three known streams beats a model turn. A media-publisher clause in the brief is the cheap improvement.